AI & Machine Learning Model for Step Data Analysis
1. Integration with Firmware & App the model into a lightweight format like:
TensorFlow Lite (TFLite) – optimized for edge devices.
ONNX (Open Neural Network Exchange) – for cross-platform compatibility.
After conversion, the model can be embedded in the firmware for on-device inference, or accessed via cloud-
based APIs if hardware constraints exist.
Real-time Inference Options:
Edge-based inference: The wearable device processes data locally using the TFLite model and
instantly detects anomalies.
Cloud-based inference: The app sends step data to the cloud (via Firebase, AWS API Gateway +
Lambda). The model runs in the cloud and returns insights to the app.
App Integration:
The Marshee app can display results like:
Your pet was very active today.
Unusual low movement detected.
Long rest periods observed.
These insights are updated in near real-time based on incoming data.
Why do we need to convert the model to something like TFLite or ONNX?
Because regular ML models (like RandomForest or full TensorFlow models) are too heavy or slow for small,
low-power devices like a pet tracker.
Think of it like this:
Full-size model TFLite / ONNX
Big & slow Tiny & fast
Needs a full PC or server Can run on a small chip or embedded system
Not optimized for low battery Designed for efficient battery + memory use
What happens after conversion?
You get a lightweight model file (.tflite or .onnx) that can be loaded by the firmware inside the
wearable.
It will let the pet tracker make real-time decisions, like example:
o “Your dog’s resting too long”
o “This walk seems shorter than usual”
Why not just keep using the model as-is?
Because:
Devices like Marshee Smart Tracker may not support full Python/Scikit-learn.
We want fast, battery-friendly decisions and no cloud delay.
It also makes offline functionality possible.
So basically, The model conversion = make it fit the device.
2. Deployment Strategy & Real-Time Inference
Over-the-Air (OTA) Model Updates:
The wearable firmware checks periodically (e.g., daily or weekly) for model updates from a secure
server.
If a new model is available, it's downloaded and replaces the old version on the device.
Versioning ensures rollback if errors occur.
Periodic Retraining:
Step data from all users is anonymized and stored in a central database.
Retraining is scheduled monthly or based on performance metrics.
After retraining, the new model is tested, converted, and pushed via OTA.
Hybrid Inference Model (Edge + Cloud)
Use a basic model on the device (for quick checks).
Use a more complex model in the cloud for deeper analysis.
The device can switch between them depending on:
o Battery level
o Internet availability
o Size of the data
3. Security & Optimization Considerations
Data Security:
Encryption: All step data is encrypted in transit (TLS) and at rest (AES-256).
User privacy: Data is anonymized before storage and model training.
Secure OTA updates: Firmware checks digital signatures before applying model updates.
Model Optimization:
Use quantized TFLite models to reduce memory usage and inference time.
Target less than 1MB model size for compatibility with low-power microcontrollers.
4. Potential AI-Driven Features
1. Anomaly Alerts
o Notify users in real time if the pet shows unusual inactivity
e.g., no steps for 3+ hours during usual active time
or hyperactivity (e.g., pacing or running unusually).
o Useful for early signs of stress, illness, or anxiety.
2. Fitness Goals
o AI recommends daily step targets based on the pet's:
Breed
Age
Historical activity levels
o Can adjust goals automatically if the pet is improving or underperforming.
3. Heatmaps of Activity
o Show active vs. resting patterns as heatmaps by:
Hour of day (circadian rhythm)
Day of week (e.g., more walks on weekends)
o Helps owners adjust routines or notice behavioral shifts.
4. Behavior Profiling
o ML can learn the pet’s baseline behavior pattern over weeks.
o Alerts are sent if deviations occur, like:
Sleeping too much
Sudden drop in walking duration
o Great for long-term health tracking.
5. Breed-Specific Activity Insights
o Compare a pet’s stats with average data from other similar pets
o E.g., “Your Beagle walks 10% less than other Beagles”.
o Helps pet parents know what’s “normal” for their pet type.
System Architecture: